Translate a SAPIEN ``PhysxCollisionShape`` into a ``ShapeSpec``.

State trajectory from the SHIPPED maniskill.<key>_native task (standard handler path).

Spherical linear interpolation between two quaternions.

Drop per-frame RGB/depth (re-renderable) -- keep the trajectory + object assets.

A smooth, demo-like action sweep that visibly moves the arm + gripper.

Computes the derivative of a sequence of SO3 rotations. Args: rotations: shape (B, 4). dt: time step. Return

Solves IK for the given action end-effector action, in either delta or absolute control

Stack the last n_obs_steps of one obs key -> (1, To, *feat), front-padded.

Copy each object's MJCF into local_pack_box/<name>/ alongside its USD.

Copy openarm_wuji robot files + meshes; return the staged MJCF path.

A list[DictEnvState] shaped exactly like state_tensor_to_nested output.

ManiSkill normalized action → absolute joint targets in the handler's sorted order.

Timeout flags. Note that max_episode_steps is set to -1 by default (no timeout).

Coerce a per-frame camera tensor into a ``(H, W, 3)`` uint-able array. Backends differ slightly: mujoco returns ``(H, W, 3)`` after the ``[0]

Track failure rate per motion bin. A frame is "failed" when anchor position error > threshold. ``_bin_failed_count`` aggregates this

Update metrics for logging.

Update contact state on the Newton path via the handler's contact reader. Reads per-env per-foot contact force from the mujoco_warp contact

Flat-terrain foot height = foot body z (ground at z=0). No ray-cast.

Bake a URDF-articulation object (at rest config) into one textured GLB. sapien3's articulation loader renders the mobility.urdf geometry but drop

Validates a (low, high) tuple used in scale-based randomization. This function ensures the tuple follows expected rules when applying a 'scale'

Validates a (low, high) tuple used in scale-based randomization. This function ensures the tuple follows expected rules when applying a 'scale'

Fail loud if the recorded ManiSkill layout differs from what we slice. Guards the hardcoded agent keys and qpos slice so a schema change raises a

A dm_control scene option that renders only VISUAL geoms (MuJoCo group 1). LIBERO/robosuite MJCFs carry both visual geoms (group 1, textured/mate

Replace the robot's free root joint with a weld to the world. Used by the `_pendant` scenes — without an RL policy, a free-floating humanoid

Family-specific one-time wiring (friction materials, checkers, object handles).

Write a minimal ManiSkill-layout ``.h5`` and return the qpos we stored. The articulation row layout mirrors ManiSkill's ``hstack([pose.p(3),

Cube + static table as scene objects for the Newton (RobotCfg) path. Mirrors ``patch_mjcf_add_cube_and_table`` geometry: a 5cm dynamic cube o

Apply both PD actuator injection AND cube/table injection to YAM.

Replay the benchmark's recorded ACTIONS and reproduce them on the handler. We rebuild a *fresh* robosuite 1.5.2 env (so its model has the current

Converts action tensor to dict with joint keys

Adaptive L2 loss.

Add materials to the MJCF asset from the URDF link.

Add appropriate lighting to the scene based on object bounds.

Set the global SAPIEN PhysX config exactly like ``BaseEnv._set_scene_config``.

Set a uniform contact material on every collision shape of a loaded rigid-body object. Mirrors a ManiSkill task that builds an object with a cust

Apply a fixed transform profile to a hand pose in XYZ+XYZW form.

Ensure that array or tensor @x has size @size in dim @dim. Args: x (np.ndarray or torch.Tensor): input array or tensor size

Axis-angle (axis*angle) -> quaternion [x,y,z,w].

One-time world bootstrap: ground plane + a default world.